Anthropic Releases Anvil, an Automated Red-Teaming System

On November 13, 2026, Anthropic formally unveiled Anvil, an automated red-teaming system built on Claude. According to Anthropic, the system is designed to autonomously probe for vulnerabilities in a model's safety defenses, with testing scoped to three areas: biosecurity, cybersecurity, and information security. Anthropic stated that Anvil's performance across multiple evaluations has surpassed that of human red-team testers, and that the time needed to uncover some vulnerabilities has dropped from weeks to hours.

After the announcement, Javier Rando, a safety and AI safety researcher at Anthropic, posted about the work on X, acknowledging the contributions of the first three co-authors named in Anthropic's technical blog post: Nate Rahn, Chawin Sitawarin, and David Huang. The results Anthropic has published show that, in biosecurity testing, Anvil located more severe jailbreak approaches in 23 of 24 test combinations, with the remaining one matching the level reached by human testers. In cybersecurity testing, Anvil successfully broke through the safety protections of all 24 task and model combinations, completing the work in 0.5 to 3.5 hours. In information security testing, Anvil managed the task while sending only 435 requests. The available reporting covers these published test outcomes only, and does not describe any further deployment plans or availability details for Anvil.